Building on Claude: what's yours to answer for.

Building on Claude raises the same two questions as any foundation model: does using it make you a GPAI model provider under Article 53, and what are you the provider of in your own right? The answers are the same as for GPT or Mistral - the Act is model-agnostic - but the specifics of how teams use Claude change which parts bite.

Disclosure: Reglynn's own reports are generated by an AI system running on Anthropic's Claude, and delivered automatically. We classify ourselves as a Limited Risk system under Article 50 and publish our own assessment. This page describes the same analysis we applied to ourselves.

Calling the API keeps Article 53 with Anthropic

Anthropic is the GPAI model provider for Claude. The Article 53 obligations - technical documentation, copyright policy, the public training-data summary, cooperation with the AI Office - sit with them, not with you, when you call the API, prompt it, or fine-tune lightly within your domain.

What would move that line is substantial retraining that materially alters the model's general capability, or repurposing it into a new task domain. Prompting, retrieval augmentation and system-prompt engineering do not. The four factors that decide it.

You are still the provider of what you ship

Claude being Anthropic's doesn't change who is responsible for your product. Ship a feature under your own name and you are the provider of that AI system under Article 3(3) - classified by what it does, not by whose model runs it.

Two things specific to how Claude tends to get used

Long-context document analysis. A common Claude pattern is feeding it large documents - contracts, CVs, case files - and asking for analysis. The classification hinges on one question: is a natural person the subject of the evaluation? Summarising a contract is fine. Scoring the candidate whose CV you pasted is profiling, and profiling closes the Article 6(3) exemption absolutely.

Tool use and agents. Where a Claude-powered agent takes actions rather than producing text, the Act still classifies it by intended purpose and effect. What changes is Article 14: if the system does land in high-risk, human oversight has to be designed so a person can actually intervene - which is harder to retrofit onto an autonomous loop than onto a chat interface. AI agents and the EU AI Act covers agents in full.

What you owe today

• Article 50 - one sentence at first contact if it converses with users. Live since 2 August 2026.
• Article 4 - recorded AI literacy for staff operating the system. What counts.
• Article 5 - no emotion inference in employment or education contexts.
• A dated written note of which systems you assessed and what you concluded - including the ones you decided were minimal risk, and why.

Anthropic publishes its own AI Act documentation for the model layer. That covers Claude; it does not cover the system you built on top, which is the part you're accountable for.

Frequently asked questions

Does building on Claude make me a GPAI model provider?

No, in ordinary use. Anthropic is the GPAI model provider for Claude, so Article 53 obligations - technical documentation, copyright policy, the public training-data summary - sit with them when you call the API, prompt it, or fine-tune lightly in your domain. Substantial retraining that materially alters general capability can change that.

What am I responsible for if I build on Claude?

The AI system you ship. Under Article 3(3) you are its provider, and it is classified by its intended purpose rather than by whose model runs underneath. In practice that means Article 50 disclosure if it converses with users, Article 4 literacy for your staff, and Chapter III only if what you built is genuinely high-risk.

Is using Claude to analyse documents high-risk?

It depends on whether a natural person is the subject of the evaluation. Summarising a contract is not. Scoring the candidate whose CV you pasted is profiling, which is high-risk under Annex III and closes the Article 6(3) exemption absolutely.

How does the AI Act treat Claude-powered agents?

By intended purpose and effect, not by degree of autonomy. Where an agentic system does land in high-risk, Article 14 human oversight becomes the demanding part - a person has to be able to intervene meaningfully, which is harder to retrofit onto an autonomous loop than onto a chat interface.

Which AI model does Reglynn use?

Reglynn reports are generated by an AI system built on Anthropic's Claude, running Reglynn's own question set and risk logic, and delivered automatically. Reglynn classifies itself as a Limited Risk AI system under Article 50 of the EU AI Act and discloses AI involvement wherever it occurs, including in its chat assistant. Our own self-assessment is published at reglynn.eu/our-ai-act-status.